Design dimensions of object-oriented languages

intro, paradigms, comparison, design, prototypes, architectures, summary, Q/A, literature

Subsections:

Despites the widespread adoption of object-oriented terminology in the various areas of computer science and software development practice, there is considerable confusion about the precise meaning of the terms employed and the (true) nature of object-oriented computing. In an attempt to resolve this confusion, [Wegner87] (in the landmark paper Dimensions of object-based language design) introduces the distinction between object-based and object-oriented. See slide 5-object-based. This distinction comes down to, roughly, the distinction between languages providing only encapsulation (object-based) or encapsulation plus inheritance (object-oriented). See section object-based. Another issue in the debate about object-orientation is the relation between classes and types. [Wegner87] concludes that the notions of objects, classes and inheritance (that constitute the classical object model) are highly interrelated, and instead proposes an orthogonal approach by outlining the various dimensions along which to design an object-oriented language. These dimensions may be characterized by the phrases: objects, types, delegation and abstraction.

Object-oriented language design

object: state + operations
class: template for object creation
inheritance: super/base and subclasses


  object-oriented = 
        objects + classes + inheritance

data abstraction -- state accessible by operations

strong typing -- compile time checking

slide: Object-based versus object-oriented

In this section we will look at the arguments presented in [Wegner87] in somewhat more detail. Also, we will look at the viability of combining seemingly disparate paradigms (such as the logic programming paradigm) with the object-oriented language paradigm. In the sections that follow, we will discuss some alternatives and extensions to the object model.

Object-based versus object-oriented

Instructor's Guide

intro, paradigms, comparison, design, prototypes, architectures, summary, Q/A, literature

How would you characterize Ada83? See [Barnes94]. Is Ada object-oriented? And Modula-2? See [Wirth83]. The answer is no and no. And Ada9X and Modula-3? See [Barnes94] and [Card89]. The answer is yes and yes. In the past there has been some confusion as to when to characterize a language as object-oriented. For example, [Booch86] characterizes Ada as object-oriented and motivates this by saying that Ada can be used as an implementation language in an object-oriented approach to program development. Clearly, Ada supports some notion of objects (which are defined as packages). However, although Ada supports objects and generic descriptions of objects (by generic packages), it does not support code sharing by inheritance. In a later work, [Booch91] revises his original (faulty) opinion, in response to [Wegner87], who proposed considering inheritance as an essential characteristic of object orientation. Similarly, despite the support that Modula-2 offers for defining (object-like) abstract data types, consisting of an interface specification and an implementation (which may be hidden), Modula-2 does not support the creation of derived (sub)types that share the behavior of their base (super)type. See also section adt-modules.

Classes versus types

Another confusion that frequently arises is due to the ill-defined relationship between the notion of a class and the notion of a type. The notion of types is already familiar from procedural programming languages such as Pascal and (in an ill-famed way) from C. The type of variables and functions may be profitably used to check for (syntactical) errors. Strong static type checking may prevent errors ranging from simple typos to using undefined (or wrongly defined) functions. The notion of a class originally has a more operational meaning. Operationally, a class is a template for object creation. In other words, a class is a description of the collection of its instances, that is the objects that are created using the class description as a recipe. Related to this notion of a class, inheritance was originally defined as a means to share (parts of) a description. Sharing by (inheritance) derivation is, pragmatically, very convenient. It provides a more controlled way of code sharing than, for example, the use of macros and file inclusion (as were popular in the C community). Since [Wegner87] published his original analysis of the dimensions of object-oriented language design, the phrase object-oriented has been commonly understood as involving objects, classes and inheritance. This is the traditional object model as embodied by Smalltalk and, to a large extent, by Eiffel, C++ and Java. However, unlike Smalltalk, both Eiffel and C++ have also been strongly influenced by the abstract data type approach to programming. Consequently, in Eiffel and C++ classes have been identified with types and derivation by inheritance with subtyping. Unfortunately, derivation by inheritance need not necessarily result in the creation of proper subtypes, that is classes whose instances conform to the behavior specified by their base class. In effect, derived classes may be only distantly related to their base classes when inheritance is only used as a code sharing device. For example, a window manager class may inherit from a list container class (an idiom used in Meyer, 1988).

Towards an orthogonal approach -- type extensions

Instructor's Guide

intro, paradigms, comparison, design, prototypes, architectures, summary, Q/A, literature

According to [Wegner87], much of the confusion around the various features of object-oriented programming languages arises from the fact that these features are largely interdependent, as for instance the notion of object and class on the one hand, and the notion of class and inheritance on the other. To resolve this confusion, [Wegner87] proposes a more orthogonal approach to characterize the various features of object-oriented languages, according to dimensions that are to a large extent independent. See slide 5-orthogonal.

Orthogonal approach

objects -- modular computing agents
types -- expression classification
delegation -- resource sharing
abstraction -- interface specification

slide: Orthogonal dimensions

The features that constitute an object-oriented programming language in an orthogonal way are, according to [Wegner87]: objects, types, delegation and abstraction.

Objects

are in essence modular computing agents. They correspond to the need for encapsulation in design, that is the construction of modular units to which a principle of locality applies (due to combining data and operations). Object-oriented languages may, however, differ in the degree to which they support encapsulation. For example, in a distributed environment a high degree of encapsulation must be offered, prohibiting attempts to alter global variables (from within an object) or local instance variables (from without). Moreover, the runtime object support system must allow for what may best be called remote method invocation. As far as parallel activity is concerned, only a few languages provide constructs to define concurrently active objects. See section Active for a more detailed discussion. Whether objects support reactiveness, that is sufficient flexibility to respond safely to a message, depends largely upon (program) design. [Meyer88], for instance, advocates a {\em shopping list approach} to designing the interface of an object, to allow for a high degree of (temporal) independence between method calls.

Types

may be understood as a mechanism for expression classification. From this perspective, Smalltalk may be regarded as having a dynamic typing system: dynamic, in the sense that the inability to evaluate an expression will lead to a runtime error. The existence of types obviates the need to have classes, since a type may be considered as a more abstract description of the behavior of an object. Furthermore, subclasses (as may be derived through inheritance) are more safely defined as subtypes in a polymorphic type system. See section flavors. At the opposite side of the type dimension we find the statically typed languages, which allow us to determine the type of the designation of a variable at compile-time. In that case, the runtime support system need not carry any type information, except a dispatch table to locate virtual functions.

Delegation

(in its most generic sense) is a mechanism for resource sharing. As has been shown in [Lieber], delegation subsumes inheritance, since the resource sharing effected by inheritance may easily be mimicked by delegating messages to the object's ancestors by means of an appropriate dispatching mechanism. In a narrower sense, delegation is usually understood as a more dynamic mechanism that allows the redirection of control dynamically. In addition, languages supporting dynamic delegation (such as Self) do not sacrifice dynamic self-reference. This means that when the object executing a method refers to itself, the actual context will be the delegating object. See section prototypes for a more detailed discussion. In contrast, inheritance (as usually understood in the context of classes) is a far more static mechanism. Inheritance may be understood as (statically) copying the code from an ancestor class to the inheriting class (with perhaps some modifications), whereas delegation is truly dynamic in that messages are dispatched to objects that have a life-span independent of the dispatching object.

Abstraction

(although to some extent related to types) is a mechanism that may be independently applied to provide an interface specification for an object. For example, in the presence of active objects (that may execute in parallel) we may need to be able to restrict dynamically the interface of an object as specified by its type in order to maintain the object in a consistent state. Also for purely sequential objects we may impose a particular protocol of interaction (as may, for example, be expressed by a contract) to be able to guarantee correct behavior. Another important aspect of abstraction is protection. Object-oriented languages may provide (at least) two kinds of protection. First, a language may have facilities to protect the object from illegal access by a client (from without). This is effected by annotations such as private and protected. And secondly, a language may have facilities to protect the object (as it were from within) from illegal access through delegation (that is by instances of derived object classes). Most languages support the first kind of protection. Only few languages, among which are C++ and Java, support the second kind too. The independence of abstraction and typing may further be argued by pointing out that languages supporting strong typing need not enforce the use of abstract data types having a well-defined behavior.

Multi-paradigm languages -- logic

Instructor's Guide

intro, paradigms, comparison, design, prototypes, architectures, summary, Q/A, literature

Object-oriented programming has evolved as a new and strong paradigm of programming. Has it? Of the languages mentioned, only Smalltalk has what may be considered a radically new language design (and to some extent also the language Self, that we will discuss in the next section). Most of the other languages, including Eiffel, C++ (and for that matter also CLOS and Oberon), may be considered as object-oriented extensions of already existing languages or, to put it more broadly, language paradigms. Most popular are, evidently, object-oriented extensions based on procedural language paradigms, closely followed by the (Lisp-based) extensions of the functional language paradigm. Less well-known are extensions based on the logic programming paradigm, of which DLP is my favorite example. In [Wegner92], it is argued that the logic programming paradigm does not fit in with an object-oriented approach. I strongly disagree with this position. However, the arguments given in [Wegner92] to defend it are worthwhile, in that they make explicit what desiderata we may impose on object-oriented languages. Remaining within the confines of a classical object model, the basic ingredients for an object-oriented extension of any language (paradigm) are: objects, classes and inheritance. Although the exact meaning of these notions is open for discussion, language designers seem to have no difficulty in applying these concepts to extend (or design) a programming language.

Open systems

flexible (dynamic) choice of actions

(static) scalability

Dimensions of modularity

encapsulation boundary -- interface to client
distribution boundary -- visibility from within objects
concurrency boundary -- threads per object, synchronization

slide: Dimensions of modularity

According to [Wegner92], the principal argument against combining logic programming and object-oriented programming is that such a combination does not support the development of open systems without compromising the logical nature of logic programming. Openness may be considered as one of the prime goals of object orientation. See slide 5-open. A software system is said to be open if its behavior can be easily modified and extended. [Wegner92] distinguishes between two mechanisms to achieve openness; dynamically through reactiveness, and statically through modularity. Reactiveness allows a program to choose dynamically between potential actions. For sequential object-oriented languages, late binding (that is, the dispatching mechanism underlying virtual function calls) is one of the mechanisms used to effect the dynamic selection of alternatives. Concurrent object-oriented languages usually offer an additional construct, in the form of a guard or accept statement, to determine dynamically which method call to answer. In both cases, the answer depends upon the nature of the object and (especially in the latter case) the state of the object (and its willingness to answer). Openness through modularity means that a system can safely be extended by adding (statically) new components. The issue of openness in the latter sense is immediately related to the notion of scalability, that is the degree to which a particular component can be safely embedded in a larger environment and extended to include new functionality. At first sight, classes and inheritance strongly contribute to achieving such (static) openness. However, there is more to modularity than the encapsulation provided by classes only. From a modeling perspective, encapsulation (as provided by objects and classes) is the basic mechanism to define the elements or entities of a model. The declarative nature of an object-oriented approach resides exactly in the opportunity to define such entities and their relations through inheritance. However, encapsulation (as typically understood in the context of a classical object model) only provides protection from illegal access from without. As such, it is a one-sided boundary. The other side, the extent to which the outside world is visible for the object (from within), may be called the distribution boundary. Many languages, including Smalltalk and C++, violate the distribution boundary by allowing the use of (class-wide) global variables. (See also section meta.) Evidently, this may lead to problems when objects reside on distinct processors, as may be the case in distributed systems. Typically, the message passing metaphor (commonly used to characterize the interaction between objects) contains the suggestion that objects may be physically distributed (across a network of processors). Also (because of the notion of encapsulation), objects are often regarded as autonomous entities, that in principle may have independent activity. However, most of the languages mentioned do not (immediately) fulfill the additional requirements needed for actual physical distribution or parallel (multi-threaded) activity.

Object-oriented logic programming

Logic programming is often characterized as relational programming, since it allows the exhaustive exploration of a search space defined by logical relations (for instance, by backtracking as in Prolog). The advantage of logic programming, from a modeling point of view, is that it allows us to specify in a logical manner (that is by logical clauses) the relations between the entities of a particular domain. A number of efforts to combine logic programming with object-oriented features have been undertaken, among which is the development of the language Vulcan. Vulcan is based on the Concurrent Prolog language and relies on a way of implementing objects as perpetual processes. Without going into detail, the idea (originally proposed in [ST83]) is that an object may be implemented as a process defined by one or more (recursive) clauses. An object may accept messages in the form of a predicate call. The state of an object is maintained by parameters of the predicate, which are (possibly modified by the method call) passed to the recursive invocation of one of the clauses defining the object. To communicate, an object (defined as a process) waits until a client asks for the execution of a method. The clauses defining the object are then evaluated to check which one is appropriate for that particular method call. If there are multiple candidate clauses, one is selected and evaluated. The other candidate clauses are discarded. Since the clauses defining an object are recursive, after the evaluation of a method the object is ready to accept another message. The model of (object) interaction supported by Concurrent Prolog requires fine-grained concurrency, which is possible due to the side-effect free nature of logical clauses. However, to restrict the number of processes created during the evaluation of a goal, Concurrent Prolog enforces a committed choice between candidate clauses, thus throwing away alternative solutions. [Wegner92] observes, rightly I think, that the notion of committed choice is in conflict with the relational nature of logic programming. Indeed, Concurrent Prolog absolves logical completeness in the form of backtracking, to remain within the confines of the process model adopted. [Wegner92], however, goes a step further and states that reactiveness and backtracking are irreconcilable features. That these features may fruitfully be incorporated in a single language framework is demonstrated by the language DLP. However, to support backtracking and objects, a more elaborate process model is needed than the process model supported by Concurrent Prolog (which in a way identifies an object with a process). With such a model (sketched in appendix E), there seems to be no reason to be against the marriage of logic programming and object orientation.

Active objects -- synchronous Java/C++

Instructor's Guide

intro, paradigms, comparison, design, prototypes, architectures, summary, Q/A, literature

When it comes to combining objects (the building blocks in an object-oriented approach) with processes (the building blocks in parallel computing), there are three approaches conceivable. See slide 6-o-conc.

Object-based concurrency

add processes -- synchronization
multiple active objects -- rendezvous
asynchronous communication -- message buffers

slide: Objects and concurrency

One can simply add processes as an additional data type. Alternatively, one can introduce active objects, having activity of their own, or, one can employ asynchronous communication, allowing the client and server object to proceed independently.

Processes

The first, most straightforward approach, is to simply add processes as a primitive data type, allowing the creation of independent threads of processing. An example is Distributed Smalltalk (see Bennett, 1987). Another example is Java, which provides support for threads, synchronized methods and statements like wait and notify to protect re-entrant concurrent methods. The disadvantage of this approach, however, is that the programmer has full responsibility for the most difficult part of parallel programming, namely the synchronization between processes and the avoidance of common errors such as simultaneously assigning a value to a shared variable. Despite the fact that the literature, see [Andrews], abounds with primitives supporting synchronization (such as semaphores, conditional sections and monitors), such an approach is error-prone and means a heavy burden on the shoulders of the application developer.

Active objects

A second, and in my view preferable, approach is to introduce explicitly a notion of active objects. Within this approach, parallelism is introduced by having multiple, simultaneously active objects. An example of a language supporting active objects is POOL, described in [Am87]. Communication between active objects occurs by means of a (synchronous) rendezvous. To engage in a rendezvous, however, an active object must interrupt its own activity by means of an (Ada-like) accept statement (or answer statement as it is called in POOL), indicating that the object is willing to answer a message. The advantage of this approach is, clearly, that the encapsulation boundary of the object (its message interface) can conveniently be employed as a monitor-like mechanism to enforce mutual exclusion between method invocations. Despite the elegance of this solution, however, unifying objects and processes in active objects is not without problems. First, one has to decide whether to make all objects active or allow both passive and active objects. Logically, passive objects may be regarded as active objects that are eternally willing to answer every message listed in the interface description of the object. However, this generalization is not without penalty in terms of runtime efficiency. Secondly, a much more serious problem is that the message-answering semantics of active objects is distinctly different from the message-answering semantics of passive objects with respect to self-invocation. Namely, to answer a message, an active object must interrupt its own activity. Yet, if an active object (in the middle of answering a message) sends a message to itself, we have a situation of deadlock. Direct self-invocation, of course, can be easily detected, but indirect self-invocations require an analysis of the complete method invocation graph, which is generally not feasible.

Asynchronous communication

Deadlock may come about by synchronous (indirect) self-invocation. An immediate solution to this problem is provided by languages supporting asynchronous communication, which provide message buffers allowing the caller to proceed without waiting for an answer. Asynchronous message passing, however, radically deviates from the (synchronous) message passing supported by the traditional (passive) object model. This has the following consequences. First, for the programmer, it becomes impossible to know when a message will be dealt with and, consequently, when to expect an answer. Secondly, for the language implementor, allocating resources for storing incoming messages and deciding when to deal with messages waiting in a message buffer becomes a responsibility for which it is hard to find a general, yet efficient, solution. Active objects with asynchronous message passing constitute the so-called actor model, which has influenced several language designs. See [Agha].

Synchronous C++/Java

In [Petitpierre98], an extension of C++ is proposed that supports active objects, method calls by rendez vous and dynamic checks of synchronization conditions. The concurrency model supported by this language, which is called sC++, closely resembles the models supported by CCS, CSP and Ada.

An example of the declaration of an active object in sC++ is given in slide ex-active.

sC++


  active class S { 
  public: 
     m () { ... } 
  private: 
     @S () {   pseudo-constructor
           select { 
              01 -> m();  external call 
              instructions ...
           || 
              accept m;   accept internal method
              instructions ... 
           ||
              waituntil (date);  time-out
              instructions ... 
           ||
              default            default
              instructions ... 
           } 
      } 
  }

slide: Synchronization conditions in sC++

The synchronization conditions for instances of the class are specified in a select statement contained in a constructor-like method, which defines the active body of the object. Synchronization may take place in either (external) calls to another active object, internal methods that are specified as acceptable, or time-out conditions. When none of the synchronization conditions are met, a default action may take place. In addition to the synchronization conditions mentioned, a when guard-statement may occur in any of the clauses of select, to specify conditions on the state of the object or real-time constraints.

The sC++ language is implemented as an extension to the GNU C++ compiler. The sC++ runtime environment offers the possibility to validate a program by executing random walks, which is a powerful way to check the various synchronization conditions. The model of active objects supported by sC++ has also been realized as a Java library, see [Petitpierre99]. There is currently, however, no preprocessor or compiler for Java supporting synchronous active objects.

As argumented in [Petitpierre98], one of the advantages of synchronous active objects is that they allow us to do away with event-loops and callbacks. Another, perhaps more important, advantage is that the model bears a close relationship with formal models of concurrency as embodied by CCS and CSP, which opens opportunities for the verification and validation of concurrent object-oriented programs. In conclusion, in my opinion, the active object model discussed deserves to become a standard for both C++ and Java, not because it unifies the concurrency model for these languages, which is for example also done by JThreads++ described in [JThreads], but because it offers a high level of abstraction suitable for concurrent object-oriented software engineering.